Ising machines are annealing processors that can solve combinatorial optimization problems via the physical evolution of the corresponding Ising graphs. Such machines are, however, typically restricted to solving problems with certain kinds of graph topology because the spin location and connections are fixed. Here, we report a universal Ising machine that supports arbitrary Ising graph topology with reasonable hardware resources using a coarse-grained compressed sparse row method to compress and store sparse Ising graph adjacency matrices. The approach, which we term interaction-centric storage, is suitable for any kind of Ising graph and reduces the memory scaling cost. We experimentally implement the Ising machine using compute-in-memory hardware based on a 40 nm resistive random-access memory arrays. We use the machine to solve max-cut and graph colouring problems, with the latter showing a 442–1,450 factor improvement in speed and 4.1 × 10⁵–6.0 × 10⁵ factor reduction in energy consumption compared to a general-purpose graphics processing unit. We also use our Ising machine to solve a realistic electronic design automation problem—multiple patterning lithography layout decomposition—with 390–65,550 times speedup compared to the integer linear programming algorithm on a typical central processing unit.

Ising 机是退火处理器，可以通过相应 Ising 图的物理演化来解决组合优化问题。然而，由于自旋位置和连接是固定的，此类机器通常仅限于解决某些类型的图拓扑问题。在这里，我们报告了一种通用伊辛机，它支持任意伊辛图拓扑，具有合理的硬件资源，使用粗粒度压缩稀疏行方法来压缩和存储稀疏伊辛图邻接矩阵。这种方法，我们称之为以交互为中心的存储，适用于任何类型的 Ising 图，并降低了内存扩展成本。我们使用基于 40 nm 电阻式随机存取存储器阵列的内存计算硬件实验性地实现了 Ising 机。我们使用该机器来解决最大切割和图形着色问题，与通用图形处理单元相比，后者的速度提高了 442–1,450 倍，能耗降低了 4.1 × 10 ⁵ –6.0 × 10 ⁵倍。我们还使用我们的伊辛机来解决现实的电子设计自动化问题——多重图案光刻布局分解——与典型中央处理单元上的整数线性编程算法相比，速度提高了 390-65,550 倍。